Cicada Lip Sync

The hard part of a talking-head video isn't the generation, it's the mouth. You've got a character and an audio track and the two refuse to agree. The Cicada Lip Sync node is the lazy person's fix for that: drop in a video, drop in audio, and a server somewhere does the syncing for you. No local model, no VRAM, no diffusion steps. This is a paid cloud API - 蝉镜 (Cicada AI) at chanjing.cc - wearing a ComfyUI node costume, and the costume fits well enough that you can forget you're making API calls until the bill arrives.

That's the thing to get straight before you use it: nothing runs on your machine. The node uploads your video and audio to the Cicada platform, asks for a video_lip_sync task, then polls a detail endpoint every five seconds until the server reports done - up to a 30-minute ceiling - and hands you back a video_url. OpenCV reads your video's dimensions and sends them along (falling back to 1080×1920 if it can't open the file), but the actual alignment and rendering all happen off-machine. Which is both the appeal and the catch: your footage leaves your computer, and every job burns the platform's credit currency (蝉豆, "Cicada Beans"). If you'd rather keep it local, the open path is a Wan Animate / LTX-2 chain - free, but fiddly, and you'll be chaining three models to get what this returns in one call. This is the "just works and charges you" alternative.

The inputs that matter:

• video_input - a video file or path, or a video object from an upstream node. Almost anything plays.
• audio_input - expects ComfyUI's AUDIO type, so it wires straight out of Cicada Voice Clone. That's the pack's signature combo: clone a voice, then lipsync a face to it.
• model - cicada-lip-sync-pro is the default and the author's recommendation; the base model is cheaper and rougher.
• backway - what happens when the video is shorter than the audio: forward loops it, reverse plays forward then backwards. Worth setting before a long monologue.
• drive_mode - normal starts from frame one, random starts somewhere random.

Output is a single video_url (STRING), and it's time-limited. Wire it straight into Cicada Video Player, which downloads it to your output folder - do that promptly, because the URL dies on a timer.

Installing it

Same story as the other two nodes in this pack. Via ComfyUI Manager, search for chanjingAI (display name "Cicada AI"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/chanjing-ai/chanjingAI-ComfyUI.git

Then the part people trip on: copy config.example.json to config.json inside the plugin folder and fill in your app_id and secret_key from https://www.chanjing.cc/platform/api_keys. Restart ComfyUI and the nodes live under the Cicada AI category. Dependencies (requests, opencv-python, mutagen) auto-install on import; you need no GPU for any of it, just network and credits.

Where people get burned

• "config.json not found" - you haven't created it yet, or it still has the your_app_id_here placeholder (the plugin detects and rejects placeholders).
• Task fails after "success" - the platform checks your Cicada Bean balance after task creation, so the failure message ("扣费失败" / insufficient balance) shows up in the polling stage, not at submit. Top up, then rerun.
• Expired URL - treat the output as hot. Download the same run.
• Video shorter than audio - the clip ends awkwardly; set backway to forward or reverse so it loops instead of stopping dead.

Worth the setup? If you're producing talking-head content and the local multi-model chain keeps eating your afternoon, yes. Just remember you're paying per job and handing your footage to a third party - that's a trade worth making deliberately, not discovering on the invoice.